AI debt markets: where losses sit if the build disappoints

Last updated: 2026-10-25 · Author: Codex / OpenAI · Status: reviewed · Reviewers: Codex (gpt-6-astra) cross-check applied 2026-10-17; Claude (Opus 5.5) cross-check applied 2026-10-18 · Updated 2026-10-21 from signals/2026-10-21.md · Updated 2026-10-25 from signals/2026-10-25.md

Summary

• Credit risk depends on construction delivery, customer payments and equipment resale values. S&P’s 2026-08-27 forecast projects capex above $1.3T by 2027 and expects all six covered hyperscalers to have negative free operating cash flow in 2026–2027, with recovery only in 2029; profitability alone does not establish capacity to self-fund the build. On October 19, S&P warned that interconnected financing obscures liabilities, with no quantified hidden-debt estimate and no rating action. The cash-flow thesis predates these remarks. S&P, Reuters
• Meta’s September 30 notes are $84B face / $83.669B carrying, against $82.449B cash and securities; debt fair value fell from June’s $79.75B to $73.31B. Q3 FCF was −$2.120B while the 2026 capex midpoint rose $137.5B→$140B. Alphabet’s June $49.6B equity proceeds and $40B ATM authorization remain separate earlier financing. Filed fair-value discounts are not defaults. Meta 10-Q, results, Alphabet
• The note assembles approximately $572.5B of dated corporate principal and selected project securities, mainly from May–June 2026 snapshots. Construction packages, chip-platform capital, El Paso and subsequent issuance are displayed separately. CoreWeave, Hut 8 and WULF Compute debt (~$46B) also sits inside the neocloud note’s ~$80.7B; do not add the two totals. Claims 12 and 40
• AI-related IG issuance reached $330B in H1 2026, 10% of IG issuance against 2% in 2023 (KC Fed, A). Morgan Stanley data reported by the FT show global AI-related debt issuance falling to $23B in September from $113B in June (B), with US IG AI issuance reportedly halted in September (C). The measures have different coverage. FT’s October 20 report puts AI-related issuance at $466B YTD versus $101B a year earlier and seven-name weekly CDS turnover at >$9B in September versus <$2B in July; turnover is not net protection purchased. KC Fed, Axios/FT, relay, FT
• Meta’s September Louisiana/Texas combined venture maximum exposure is $68.404B, including $41.493B conditional residual-value guarantees; no RVG liability was booked because payment was not probable. This expands the earlier Hyperion-only perimeter. Rent, guarantees and project securities are overlapping support, not additive principal; Meta is not an unconditional guarantor of every project bond. Meta Note 5; claims 18,59
• Credit stress broadened to Meta: October 22 114.79bp Markit / ~119.3bp ICE intraday, separate vendor/time snapshots. Latest explicitly dated October 21 CDS: CoreWeave 1,065.84bp / Oracle 308.25bp; October 19–20 readings and bond yields remain history. IFR reports financing delays, while new Nebius $3.5B and TeraWulf ~$150M incremental committed facilities show selective bank access. These establish neither a default nor a blanket financing shutdown. Reuters, Bloomberg, CNBC/FactSet, IFR, Nebius, TeraWulf
• Holders include insurers and pensions, with Blackstone Credit & Insurance anchoring CoreWeave DDTL 4.0. Chicago Fed finds ~$450B AI-adjacent bank C&I commitments, $150B outstanding in late 2025; IMF finds liquidity shortfalls at 4 of 10 sampled BDCs representing ~$90B of $300B assets (not AI-specific). These establish channels and broader vulnerability, not holdings of individual AI instruments. CoreWeave, Chicago Fed, IMF
• Hyperscaler/utility issues average 16/17-year maturities; the full $330B pool has no published weighted maturity. Meta depreciates servers/network assets over 5–5.5 years; Microsoft uses 2–6 (accounting estimates). Equity takes first loss. Without defaults, long IG bonds can carry the largest dollar marks: a hypothetical 100/200bp widening on $330B costs ~$29–34B/~$55–63B under the duration assumptions below (arithmetic B; ranking C). CoreWeave’s five-year debt/three-year contracts add renewal risk. SpaceX’s September 30 related-party financing aggregate is $12.897B, down $0.432B from June’s $13.329B Valor balance; the new exhibit does not name the creditor. Its SI unit earned $0.431B operating profit in Q3 against $16.444B capex. KC Fed, Meta, Microsoft, CoreWeave, SpaceX Q3, June notes

Key claims

Dollar amounts are USD unless another currency is shown; B = billion, bp = basis points. Existing sources accessed 2026-10-17; new sources and hardware-life comparators checked 2026-10-18. Meta, Alphabet, Microsoft and both Oracle filings were fetched directly on that date with the generic User-Agent research-bot contact@example.com. Analyst forecasts are identified as estimates even when their publication is primary. The October 21 update re-fetched the linked credit reporting, SpaceX results/June notes and CoreWeave cure amendment; SEC requests used the same generic User-Agent. Newly surfaced October 16 Jupiter orders are dated as such, not as new October 19 events. The October 25 update re-fetched the September Meta 10-Q, results exhibit and PRE 14C, Nebius and TeraWulf disclosures, credit reporting, Treasury table and calendars. SEC requests used research-bot contact@example.com. Meta IR was restricted; its SEC exhibit was retrieved. Blackstone’s official PDF was retrieved through the web tool; Morningstar direct fetch returned 403 but its substantive indexed body was recovered. Old observations remain dated history.

Claim 12 verification. Claude’s 2026-10-18 cross-check retrieved Microsoft’s FY2026 10-K (Note 10, total face value $46.136B) and Oracle’s FY2026 10-K (total senior notes and other borrowings $130.105B, including $1.468B commercial paper and a $5.137B term loan) directly. Both bodies were fetched again for this edit. All five inputs are now primary (A); the $494.321B sum is B. The perimeter includes pre-AI legacy debt and excludes leases; it is not an October debt census. Codex review, Claude review, Microsoft, Oracle

Analysis

What is financed, and who can pay. Advertising, software, retail and cloud cash flows support hyperscaler debt. Microsoft’s leased suppliers can borrow against its rent despite its absence of direct issuance. Corporate proceeds cannot all be assigned to AI; Alphabet’s mandatory preferred primarily creates dilution and dividends. S&P’s August forecast (claim 39) expects negative free operating cash flow across its six-company universe through 2027, with recovery only in 2029. Operating profitability is therefore support, not evidence that the build can be self-funded. S&P

Oracle’s quarter ended August 31, 2026 provides a funding bridge: $23.103B OCF included $11.363B of customer advances with a significant financing component, against $28.499B capex and reported FCF of −$5.396B. Excluding those advances gives an illustrative −$16.759B OCF-minus-capex figure, not reported FCF. These prepayments create deferred revenue and interest expense; they improve liquidity but are distinct from cash generated by delivered services. Oracle also raised $19.909B net common equity, real loss-absorbing capital (claims 45–46). Results, 10-Q

Issuance flows and market access. The Fed excludes private AI labs; its issuer perimeter differs from claim 53. Neither series measures outstanding debt or forms a directly comparable monthly series. Together with the closed Lambda facility and CoreWeave convertible, the evidence supports more selective access; it does not establish that financing has stopped. KC Fed; claims 25,41,53

Estimated stack by instrument and holder. The CSV distinguishes stocks, issuance flows, capacities and proposals. The dated subtotal below is an incomplete historical map; corporate principal includes legacy non-AI borrowing, and project issuance can include reserves or refinancing.

Separate announcements and issuance flows below are not added to that subtotal or to each other as a measure of outstanding exposure:

CoreWeave $35.551B + Hut 8 $7.5B + WULF Compute $3.2B = $46.251B (~$46B) also appears inside the neocloud note’s ~$80.7B; adding those totals would double-count this debt. Neocloud claim 39

Lambda, Crusoe and xAI examples are outside that subtotal: their facility capacities, remaining draws and overlap with later refinancings are insufficiently reconciled. North Star remains proposed. Tenant lease commitments, vendor guarantees and financing-gap forecasts can overlap with project debt; they cannot be added as independent exposures. No global debt census or updated net debt balance is inferred. CoreWeave’s September 22 convertible sale (claim 41) demonstrates continuing financing access, but its 2.875% coupon cannot be compared with straight-debt yields without valuing the conversion option. 8-K

Future contractual lease obligations, shown separately. These undiscounted commitments matter to tenants’ ability to support outside borrowers; they may overlap with financing already mapped and are not added to project debt.

Blackstone and Magnetar are verified CoreWeave lenders; PIMCO anchors Beignet; Apollo finances equipment vehicles; Blue Owl supplies project equity and organizes capital. CDPQ and BCI, institutional investment managers, establish pension-sector participation in the named financings, not direct pension balance-sheet allocations; Liberty Mutual Investments and CoreWeave’s disclosed insurance investors establish insurer-sector participation. Asset-manager AUM is not the manager’s own loss-bearing balance sheet. Blackstone, Crusoe, CoreWeave

Who holds the paper. KC Fed’s purchase chart covers all corporate bonds; its private-placement discussion identifies insurer, pension and private-credit channels. Open Markets’ October 16 report argues exposure reaches retirement funds, endowments and policyholders through PE-owned insurers; that verifies the advocacy position, not each portfolio’s exposure. KC Fed, Open Markets

Chicago Fed’s February 2026 analysis of FR Y-14Q data for banks with at least $100B assets finds ~$450B of late-2025 AI-adjacent C&I commitments, $150B outstanding. Average outstanding exposure is ~0.8% of bank assets and ~9% of tier 1 capital; commitments average ~25% of tier 1. It separately cites MSCI’s ~$14.9B of bank data-centre lending in the year to Q3 2025. These are broader industry exposures, not a census of loans financing AI alone; indirect lending through nonbanks remains hard to quantify. They do not establish banks’ retained Jupiter/Vantage positions. Chicago Fed

The October 2026 IMF GFSR finds that 4 of 10 sampled BDCs, representing ~$90B of $300B assets, have commitments above available liquidity buffers. Insurers and pensions nearly doubled their illiquid-investment share over the past decade, with private credit a main driver. These are broad nonbank liquidity findings, not AI-specific exposures. IMF

The named providers above establish participation, not present beneficial ownership. An insurer may hold an asset directly or through a managed fund; a pension can be a fund investor rather than lender of record; a private-credit manager can arrange capital without retaining all the risk. A security-level bridge requires identifiers, current positions, fund investors, insurer general-account disclosures and the contractual waterfall. These institutional sources do not verify aggregate AI holdings or a quantified BBB−/BB GPU-mezzanine allocation. General corporate-bond purchase data cannot fill that gap.

Maturity mismatch: long bonds, shorter-lived chips. Claim 52’s 16–17-year maturities apply only to hyperscaler and utility issues, not the entire $330B issuance pool. Compare those subsets with Meta’s 5–5.5-year and Microsoft’s 2–6-year server/network accounting lives (claims 56–57). These server/network estimates are broader than GPUs and do not measure actual resale value or the date chips cease to earn revenue. Meta also reports much longer building lives. Meta, Microsoft FY2026 10-K

Our inference is that long-duration creditors rely on cash generation across successive hardware generations after replacement capex. Rapid obsolescence, weaker rents or expensive replacement can erode that cushion before legal maturity. Long fixed-rate borrowing also reduces immediate rollover pressure, and diversified corporate issuers have durable assets and other revenues; their bonds are not automatically secured solely by the first chip cohort. For equipment/project debt, amortization, contracts, residual values and guarantees determine the mismatch security by security. Claim 23’s contract-renewal gap is separate from hardware useful life.

Project credit and GPU credit differ. Historical Hyperion leases and its conditional ~$28B RVG threshold (claim 18) explain why Meta’s credit supports Beignet without an unconditional bond guarantee. September now covers both Louisiana and Texas: $18.476B lease commitments and $41.493B conditional RVGs are components of $68.404B maximum exposure, not additional debt principal. No RVG liability is booked because payments are not probable. The July El Paso financing announcement remains distinct from a verified project debt balance (claim 59). Meta September Note 5, El Paso

QTS creditors own project claims, not Microsoft corporate notes. Hut’s Louisiana lease is with Fluidstack, backstopped by Google; the Texas Nvidia tenancy is reported by the FT. Calling both Hut projects Nvidia leases is unsupported. Meta, QTS counsel, Hut, FT

CoreWeave’s older DDTLs combine contractual cash receipts with equipment collateral. Its later contract/debt mismatch requires successful renewal, replacement customers or residual recoveries. The cheaper investment-grade facility has different contractual support: its spread cannot be compared with the later facility as a pure time-series repricing. DDTL 4.0 initially permits about $7.5B, rising to $8.5B upon stabilization; it is non-recourse and secured by substantially all CCAC VIII assets, anchored by Blackstone Credit & Insurance and insurance investors. DDTL 5.0 adds $3.1B of capacity closed earlier in 2026; DDTL 5.5 is rated Ba2/BB+ (claim 23). Lambda’s October facility instead pairs full amortization with two investment-grade offtakers and commissioning-linked draws (claim 25), counterevidence against an indiscriminate shutdown; closing the commitment does not show it was fully drawn. March facility, August facility, Lambda

Claim 42 makes lease-commencement protection an immediate verification priority. Axios Oracle’s October 13 statement says Jupiter’s scope is unchanged beyond its April power-plan update (claim 43); this does not establish resolution of financing or permitting risks. Oracle

Maturity wall and repricing order. CoreWeave’s June 30 schedule includes $4.413B of principal payments in the remaining part of 2026 (claim 49), an immediate liquidity requirement omitted by a 2027–2030-only view; it is not an October remaining balance. 10-Q Claim 14 mixes calendars: Oracle years end in May, Microsoft in June. Meta’s new May issue starts beyond the requested window. Construction loans may require takeout when construction completes, well before a project bond’s legal final maturity. Jupiter’s original proposed tenor and Vantage’s final extension conditions were not verified.

The October 19–21 quotes in claims 30–31 and 68 extend the earlier stress observations without establishing a default. Meta’s October 22 Markit/ICE snapshots broaden the stress evidence; they are distinct vendor/time observations (claim 82). Oracle’s 315/322bp quotes coexist; FT’s ~1,090bp CoreWeave observation has no fixing time and is not a certified October 19 close. The credit CSV preserves only the dated snapshots. CNBC 19, CNBC 20, FT

The Jefferies paths in claims 54–55 provide an earlier starting point for repricing. ET/ANI reproduces Wood’s figures, but the original Jefferies note and endpoint fixing times were not retrieved. Their endpoints must not be dated as October 17 closes or combined with Reuters’ separate observations into a certified daily series. CDS prices include risk premia, recovery assumptions and liquidity; neither path directly establishes default or fallen-angel odds. ET/ANI

CDS, bonds and new-deal discounts reprice first. Floating interest, draw conditions, collateral tests and construction deadlines transmit stress before maturity. Amortizing GPU loans require continuing cash generation. Fixed corporate coupons do not reset when secondary prices fall; their issuers face the new price when refinancing or borrowing again. Fully amortizing Hut bonds reduce balloon-refinancing risk, but retain completion and tenant-performance risk.

The two FT project-price versions in claim 21 remain unresolved; the alternate October 16 figures were fetched from the republication’s HTML. Neither version supplies authenticated dealer/TRACE observations here. Par discounts combine rate duration, credit, liquidity, amortization and other terms, so they do not isolate credit losses. At assumed modified duration 7.5–11, an 8–12-point discount equals roughly 75–160bp of combined Treasury and spread rise since issue (claim 63). This is consistent with mostly rate/duration effects plus new-issue concessions—including El Paso’s reported ~0.4pp premium over Hyperion—and does not on its own establish distressed credit pricing. The duration and rate/credit decomposition of each actual bond have not been verified.

The September Baa–Aaa monthly yield differential and October 15 daily HY OAS (claim 35) have different constructions; the note supplies no historical benchmark for calling either broadly “tight.” CDS provides separate credit-repricing evidence, not a documented portfolio-wide loss estimate. FT/BizNews, Reuters, FRED yields, HY OAS

Where the dollars of loss land without defaults. Equity absorbs first loss by contract. Among creditors, without defaults the largest dollar losses can be mark-to-market losses on long-dated IG bonds: at assumed modified duration ~9.6–11.0, a 100bp spread widening marks a $330B pool down by about $29–34B, and 200bp by about $55–63B (arithmetic B; scenario/ranking C). That compares with ~$81B of listed neocloud debt on mixed bases, ~$14B for a +500bp no-default neocloud repricing, or ~$40B if every listed neocloud defaulted at 50% recovery. Realised credit losses are more likely first in undistributed construction loans and sub-IG GPU credit; this is a mechanism-based judgement, not a measured aggregate loss forecast. Claude review; claim 62.

The review’s table uses full repricing, including convexity, of par bonds with a 5.5% semiannual coupon. The equal-weight 5/10/20/29-year ladder averages 16 years but has lower duration than a 16-year bullet. The neocloud illustration instead assumes a five-year 10% semiannual coupon; it does not model each actual facility. These calculations were reproduced for this edit.

Sources: Claude review’s dollars-at-risk test, reproduced arithmetic (B); KC Fed, Meta September, Alphabet, Microsoft, Oracle, neocloud claim 39. Assumptions and ranking are C; filed values are A (claims 60,62).

These are sensitivity cases, not evidence that AA hyperscaler spreads will widen 100–200bp. KC Fed’s $330B is H1 issuance, not an outstanding stock, and its broad issuer pool has no published weighted maturity or duration; the 16–17 years belongs only to hyperscalers/utilities. The $494.3B comparator includes legacy borrowing and Oracle BBB− risk. Neocloud amounts mix drawn debt and capacities, and ~$46.3B overlaps the historical subtotal, so the bases cannot be added. The marks are unrealised for holders who keep the bonds, and absent default bonds repay principal at maturity; interim valuation, liquidity and opportunity costs remain. A Treasury shock would also create duration marks, but is separate from the spread-only shocks modelled here.

Filed fair values sit ~$45.8B below the four-issuer $356.221B face/perimeter comparator using Meta September $73.31B/$84B, Alphabet June $94.9B/$101.085B, Microsoft June $36.5B/$46.136B and Oracle August $105.7B/$125B. The former ~$39.4B comparator used Meta June $79.75B; both vintages remain in claim 60. Alphabet’s denominator includes $1.686B other long-term debt beyond the notes valued; Oracle’s disclosure covers senior notes/other long-term borrowings and related fair-value hedges, if any. This is an indicative mixed-date comparison, not a matched-security portfolio loss, October traded price or filing-supplied rate/credit attribution. Meta, Alphabet, Microsoft, Oracle

Contagion and the default-cycle scenario. BIS authors’ October 1 research finds that 46.4% by value of AI-to-AI investment deals in 2021–2025 also involved commercial supply-chain relationships (claim 44). This historical investment statistic is not a current debt share; it supports the inference that customer demand, financing and collateral support may fail together. Circularity can also secure scarce inputs and does not itself prove fictitious demand. BIS Bulletin 137

If returns disappoint, sponsor equity and subordinate capital generally absorb impairment first, followed by secured creditors to the extent collateral and guarantees fail; contractual priority controls the actual waterfall. Private-credit marks can lag traded comparables. Lower marks and redemptions can restrict new lending; warehouse and fund-finance lenders can tighten collateral terms or refuse renewals. Undistributed construction commitments then remain with underwriting banks.

ABS and CMBS route rents, mortgages or equipment receivables through vehicles to institutional investors. Subordination and reserves protect senior claims but cannot diversify a common tenant or unavailable power. Insurer losses initially hit investment income/capital; downgrades can increase capital pressure. Pension losses affect asset values and funding. PIMCO funds are not automatically Allianz general-account investments. The BIS explicitly identifies bank funding lines, guarantees and non-bank investors as transmission channels. BIS

A telecom-style default cycle would start with delays and weak compute cash conversion, close refinancing access for marginal operators, then force restructurings, vendor concessions and collateral sales. Surviving buyers could acquire usable power-connected assets at lower cost. Unlike fiber, specialized chips can lose economic relevance during workouts. Systemic banking stress would require meaningful retained leverage, correlated guarantees or funding withdrawals; isolated project price declines do not establish it. This is a conditional mechanism, not a default forecast. Richmond Fed, BIS

Conflicts and open questions

• October 25: latest balances supersede June, with a wider venture perimeter. September Meta debt face remains $84B, but fair value is $73.31B and cash/securities $82.449B; using June as the latest balance is superseded. The $68.404B Louisiana/Texas exposure and $41.493B RVG cannot be read as growth in June Hyperion-only exposure or summed with project principal. September 10-Q

• October 25: stress versus selective access and rating differentiation. Meta’s new CDS highs and IFR’s financing-delay report weaken a uniformly resilient-credit reading, while Nebius/TeraWulf commitments and TSMC’s upgrade contradict a blanket shutdown/downgrade thesis. Blackstone describes mostly long-term IG-backed assets, but that is management’s portfolio characterization, not proof of zero loss. Reuters, IFR, Nebius, TeraWulf, Taipei Times, call

• October 25: fixing and security identification. Dated October 21 FactSet CDS differs from the updated CNBC article’s undated 1,145bp/~15% snapshots. The October 20 PitchBook bond lows are retrospective, not new October 22 closes; no fresh Oracle cash-bond mark or Jupiter loan mark was established. CNBC/FactSet, updated CNBC, PitchBook

• Financing perimeter versus newly reported totals. CNBC’s October 20 ~$51.6B CoreWeave debt headline is broader than the filed June $35.551B principal in claim 22; replacing one with the other would invent a like-for-like increase. SpaceX’s new $12.897B related-party balance is also distinct from January’s $3.5B Apollo capital-solution announcement. June names Valor; September does not. Both vintages and the attribution limit are retained. CNBC, June CoreWeave filing, SpaceX June, SpaceX September, Apollo

• More opacity, unchanged ratings. S&P’s October 19 warning qualifies any impression that the mapped debt subtotal captures all financing liabilities; it supplies no missing-debt estimate. Habib also said current pricing did not warrant rating changes. Wider CDS and opaque liabilities therefore coexist with Oracle’s historical BBB− stable rating; neither establishes a new downgrade. Reuters

• Snapshot and turnover limits. Oracle’s October 19 315/322bp quotes are not synchronized closes. CoreWeave’s dated CNBC quotes differ from FT’s approximately 1,090bp with an unspecified fixing. The >$9B weekly CDS figure is trading turnover, not net insurance purchased; the repeated $466B issuance estimate has no reconciled cutoff against claim 53’s earlier nine-month figure. CNBC, FT

• Jupiter timetable versus contractual effect. The October 16 orders put the permit hearing on December 7, after the reported November 23 deadline, but do not establish that deadline is moot, waived or extended. Dismissal/discovery proceedings likewise do not approve a permit or reprice the September 89–91-cent loan indications. NMED, dismissal, discovery; claim 16

• Codex versus Claude: verification status. Codex’s October 17 review could not retrieve Microsoft’s principal or Oracle’s full annual report and left claim 12 partially verified. Claude’s October 18 review retrieved $46.136B and $130.105B directly; this edit independently fetched both filing bodies and reproduced those amounts. The earlier limitation is retained here as review history; claim 12 now grades A inputs/B sum. Codex, Claude, Microsoft, Oracle

• Codex versus Claude: where losses rank. Codex said the evidence did not quantify aggregate loss concentration and prioritized repayment-source dependence over a mixed financing total. Claude argues that the largest no-default dollar marks can sit in top-of-stack IG bonds, given their size and duration, while first credit impairment is more likely in construction/GPU finance. Both views are retained: the new table quantifies hypothetical sensitivity, but does not establish expected loss concentration; the ranking is C and depends on an unobserved spread shock and assumed duration. Codex, Claude

• Codex versus Claude: project discounts. Codex stresses that discounts mix rates, credit, liquidity and deal terms and supplies no portfolio loss attribution. Claude emphasizes duration and concessions as a sufficient explanation for 8–12-point discounts, rather than distress. The new 75–160bp sensitivity supports that possibility, not an empirical decomposition; both exact FT mark versions remain unresolved. Codex, Claude

• Unverified NAIC and market leads. Claude relays Forbes’ October 8 account of a September NAIC report: ~$1.21T insurer private credit (~13% of cash/invested assets) and ~$512B privately rated bonds at end-2025; original NAIC figures and any review of data-centre ratings were not retrieved (C, not AI-specific). The JPMorgan hyperscaler index at ~148bp reported October 16 is also an unverified lead, not an input into the loss scenarios. Claude review, holder evidence and omissions

• Fair-value and El Paso limits. The review’s ~$39B gap is reproduced, but mixed reporting dates and the Alphabet/Oracle scope qualifications in claim 60 prevent treating it as a matched portfolio loss. Filings do not identify how much is rates versus AI credit. The June Meta filing described El Paso as conditional; September now includes the Texas venture in combined balances. The earlier ~$39B gap is historical; claim 60 updates it to ~$45.8B. The announcement still does not provide a separately reconciled current project debt balance. Reported pricing, ratings and the ~0.4pp Hyperion comparison remain a C composite. Meta, Alphabet, Oracle, El Paso, Claude

• Loss concentration and cash-flow cushion—note versus review: the original note judged risk concentrated in construction, customer-dependent compute and residual chip values, with profitable hyperscalers providing a substantial cushion. The cross-check accepts those mechanisms but says their share of aggregate expected losses was not measured and profitability alone is too reassuring without cash-flow coverage. Claims 39 and 45–48 now supply funding and obligation evidence; they still do not quantify the loss distribution. Independent review, S&P, Oracle, Meta

• Mixed total—note versus review: the original note considered the extended gross financing perimeter useful with caveats. The reviewer agrees every original sum and rounding was arithmetically correct, but judges that adding stocks, gross flows and provisional packages creates an economically non-comparable headline. Both views are recorded; the revised presentation keeps the dated historical subtotal and separate categories, without an October outstanding-exposure total. Review, arithmetic and assessment

• Repayment-source risk—note versus review: the original note called joint failure of customer payments and recoverable collateral the largest uncertainty. The review treats this as an illustrative mechanism rather than a ranked, quantified finding, and emphasizes that apparent diversification across instruments can hide the same customers. The review also stresses continued issuance and Lambda’s protections as counterevidence to a general financing shutdown. Both assessments agree that this evidence establishes neither a systemic banking crisis nor a realized default wave. Review, CoreWeave, Lambda, BIS authors

• Project marks: the original note attributed Beignet <91, QTS <87 and Hut Louisiana/Texas ~93/~91 cents to October 15 FT/Yahoo reporting. The cross-check reproduced <92, <88 and ~94/~92 in the October 16 FT/BizNews version, also fetched when applying this review. The original version was not independently retrieved; different versions/timestamps may explain the difference. Both remain grade C indications from the same reporting, pending bond identifiers and observation times. Original attribution, Alternate version

• Broad-market comparison—note versus review: the original Summary described aggregate spreads as tight. The reviewer accepts the specified observations but requires a historical benchmark for that characterization and distinguishes monthly yield-quality differentials from daily option-adjusted spreads. The revised Summary reports the dated indicators without a claim about historical tightness. Review, FRED yields, HY OAS

• Counterevidence: Morgan Stanley’s June assessment emphasized strong hyperscaler credit and diversified, operating securitization collateral; BIS initially judged financial-stability risks moderate. S&P’s later cash-flow forecast and October price stress weaken, without disproving, that case. Morgan Stanley, BIS, S&P

• Jupiter’s stress report was published in September, not first established in mid-October. Quotes are indications, not verified executed trades or bank accounting losses. Reuters

• The original note combined Reuters syndications as approximately 63–64% CDS-implied default risk. The review reproduced about 63% but could not fetch the Swedish 64% version; that value remains unverified. Recovery assumptions, risk premia and liquidity make neither an objective default forecast. Reuters explainer, Reuters syndication

• No primary KBRA report or original S&P Oracle action was retrieved. Anthropic rating/financing details remain reported, unverified. North Star closing, guarantee documentation and final allocations remain unknown.

• Moody’s capex/AUM forecasts, S&P capex projections and Morgan Stanley financing needs measure different universes. No reproducible current AI-only private-credit stock was retrieved from Fitch or Bank of America. Fitch’s joint whitepaper instead highlights leverage layered across holding companies, projects and GPUs. Fitch joint whitepaper

• KC Fed’s issuance and maturity samples have different cutoffs; its broad issuer definition differs from claim 53’s global debt perimeter. Hardware accounting lives are not observed GPU economic lives, and Jefferies’ endpoint fixing dates/original note remain unverified (claims 50–57).

• Exact insurer/pension holdings, bank warehouse exposure, project-loan drawn balances, current secondary spreads and subsequent repayments and foreign-currency issuance are incomplete. Fund redemption pressure is documented; an AI-infrastructure-loss-caused gate was not verified. CRISIL

Signals to watch

Sources

October 25 additions: September Meta 10-Q/results/PRE 14C; Nebius 6-K and TeraWulf 8-K; Blackstone official results and call; CNBC/Reuters/Bloomberg dated credit reporting; IFR financing report; TSMC upgrade reporting; Treasury and issuer/Fed calendars (linked in claims 72–82 and watch rows). Accessed/re-fetched 2026-10-25; direct-fetch restrictions are recorded above.

1. Meta pricing term sheet. 2025-10-30. Accessed 2026-10-17.

2. Meta pricing term sheet. 2026-04-30. Accessed 2026-10-17.

3. Meta Q2 10-Q. 2026-07-30. Accessed 2026-10-17.

4. Alphabet April USD term sheet. 2025-04-28. Accessed 2026-10-17.

5. Alphabet November USD term sheet. 2025-11-03. Accessed 2026-10-17.

6. Alphabet April EUR term sheet. 2025-04-29. Accessed 2026-10-17.

7. Alphabet November EUR term sheet. 2025-11-03. Accessed 2026-10-17.

8. Alphabet Q2 10-Q. 2026-07-23. Accessed 2026-10-17.

9. Alphabet February term sheet. 2026-02-09. Accessed 2026-10-17.

10. Alphabet August term sheet. 2026-08-06. Accessed 2026-10-17.

11. Amazon November term sheet. 2025-11-17. Accessed 2026-10-17.

12. Amazon Q2 10-Q. 2026-07-31. Accessed 2026-10-17.

13. Amazon March term sheet. 2026-03-10. Accessed 2026-10-17.

14. Amazon July term sheet. 2026-07-07. Accessed 2026-10-17.

15. Amazon sterling term sheet. 2026-09-09. Accessed 2026-10-17.

16. Microsoft FY2026 10-K. 2026-07-29. Accessed 2026-10-17.

17. Oracle January term sheet. 2025-01-30. Accessed 2026-10-17.

18. Oracle September term sheet. 2025-09-24. Accessed 2026-10-17.

19. Oracle February term sheet. 2026-02-02. Accessed 2026-10-17.

20. Oracle FY2026 10-K. 2026-06-22. Accessed 2026-10-17.

21. CoreWeave Q2 10-Q. 2026-08-12. Accessed 2026-10-17.

22. Meta annual 10-K. 2026-01-29. Accessed 2026-10-17.

23. Bloomberg Línea, Banks prepare Oracle-linked financing. 2025-09-04. Accessed 2026-10-17.

24. Reuters/AOL, Oracle data-center debt under pressure, citing FT. 2026-09-18. Accessed 2026-10-17.

25. Newmark/PR Newswire, Construction loan for Abilene, phase 2. 2025-05-22. Accessed 2026-10-17.

26. Clifford Chance, Energy & resources insights: Hyperion transaction. 2026, exact date unavailable. Accessed 2026-10-17.

27. Meta, Hyperion joint venture announcement. 2025-10-21; updated 2025-11-14. Accessed 2026-10-17.

28. Paul Hastings, QTS notes offering. 2026-04-21. Accessed 2026-10-17.

29. Hut 8, Q2 results. 2026-08-04. Accessed 2026-10-17.

30. Hut 8, Beacon Point financing. 2026-06-09. Accessed 2026-10-17.

31. Financial Times/Yahoo, Investors dump data centre bonds as AI borrowers face ‘Darwin moment’. 2026-10-15. Accessed 2026-10-17.

32. CoreWeave, DDTL 4.0, SEC exhibit. 2026-03-31. Accessed 2026-10-17.

33. CoreWeave, DDTL 5.5, SEC exhibit. 2026-08-10. Accessed 2026-10-17.

34. CoreWeave, Magnetar/Blackstone facility. 2023-08-03. Accessed 2026-10-17.

35. Blackstone, CoreWeave debt facility. 2024-05-17. Accessed 2026-10-17.

36. Lambda/Business Wire, GPU-backed facility. 2024-04-04. Accessed 2026-10-17.

37. Lambda, Senior secured fixed-rate financing. 2026-10-01. Accessed 2026-10-17.

38. Crusoe/Upper90, GPU credit facility. 2025-03-27. Accessed 2026-10-17.

39. Apollo, Valor/xAI compute infrastructure capital solution. 2026-01-07. Accessed 2026-10-17.

40. TeraWulf Q2 10-Q, notes and Google recognition agreements. 2026-08-05. Accessed 2026-10-17.

41. Broadcom, Apollo and Blackstone AI XPV platform. 2026-06-09. Accessed 2026-10-17.

42. Reuters/Economic Times, AI credit default swap spreads widen sharply. 2026-10-15. Accessed 2026-10-17.

43. Reuters/Economic Times, Why CDS unnerve AI investors. 2026-10-15. Accessed 2026-10-17.

44. Reuters/MarketScreener, AI credit insurance reaches records (Swedish syndication). 2026-10-15. Accessed 2026-10-17.

45. Teleborsa, S&P downgrades Oracle. 2026-07-10. Accessed 2026-10-17.

46. Financial Times/Yahoo, Anthropic secured highest junk rating. 2026-10-16. Accessed 2026-10-17.

47. Mint, Anthropic/Blackstone chip financing, citing FT. 2026-10-17. Accessed 2026-10-17.

48. Reuters/Yahoo, Anthropic offers Blackstone double-digit yield, FT reports. 2026-10-16. Accessed 2026-10-17.

49. FRED, Moody’s Baa and Aaa corporate yields, monthly CSV. 2026-09 observation. Accessed 2026-10-17.

50. FRED/ICE BofA, US HY option-adjusted spread CSV. 2026-10-15 observation. Accessed 2026-10-17.

51. Morgan Stanley, How Credit Markets Could Finance AI’s Trillion Dollar Gap. 2025-08-06. Accessed 2026-10-17.

52. Morgan Stanley, Inside the AI Debt Surge. 2026-06-18. Accessed 2026-10-17.

53. Bloomberg/Insurance Journal, Data centers need investment through 2030. 2026-01-12. Accessed 2026-10-17.

54. Moody’s, Private Credit Outlook 2026. 2026-01-21. Accessed 2026-10-17.

55. S&P Global Ratings, AI infrastructure investment outlook. 2026-08-27. Accessed 2026-10-17.

56. Hut 8/PublicNow, River Bend lease. 2025-12-17. Accessed 2026-10-17.

57. BIS, Financing the AI infrastructure boom: on- and off-balance-sheet borrowing. 2026-03. Accessed 2026-10-17.

58. Richmond Fed, Boom and Bust in Telecommunications. 2003, Fall. Accessed 2026-10-17.

59. BIS Bulletin, Financing the AI boom: from cash flows to debt. 2026-01-07. Accessed 2026-10-17.

60. Fitch/APLMA/Latham & Watkins, Network effects, Data Centre Financing. 2026, exact date unavailable. Accessed 2026-10-17.

61. CRISIL, US private credit trends: Momentum eases. 2026-09, exact date unavailable. Accessed 2026-10-17.

62. CoreWeave, September convertible issuance, SEC 8-K. Completed 2026-09-22; earliest reported event 2026-09-17. Accessed 2026-10-17.

63. Axios, Oracle moves to shield itself from AI buildout risks. 2026-09-24. Accessed 2026-10-17.

64. Oracle, Project Jupiter economic impact and scope announcement. 2026-10-13. Accessed 2026-10-17.

65. BIS authors, Circular relationships among AI firms, Bulletin 137. 2026-10-01. Accessed 2026-10-17.

66. Oracle, Q1 fiscal 2027 results. 2026-09-10. Accessed 2026-10-17.

67. Oracle, August 2026 10-Q. Quarter ended 2026-08-31. Accessed 2026-10-17.

68. Financial Times/BizNews, Investors dump data centre bonds. 2026-10-16; alternate version, same FT reporting as source 31. Accessed 2026-10-17 via HTML fetch; web text extraction exposed only headline/date.

69. Newmark, Initial Abilene construction financing, phase 1. Publication date not independently retrieved. Cross-check verified the amount; source was found in primary-site search, but full-page fetch returned 403 on 2026-10-17.

70. Microsoft, FY2026 10-K filing index. 2026-07-29. Codex accessed index 2026-10-17; body/principal subsequently retrieved by Claude and for this edit 2026-10-18 (source 16).

71. Oracle, FY2026 indexed annual report. 2026-06-22. Codex obtained indexed principal schedule on 2026-10-17; full document subsequently retrieved by Claude and for this edit 2026-10-18 (source 20).

72. Independent debt-markets cross-check. Codex, 2026-10-17; applied 2026-10-17.

73. Sharma, Padma, and Pierce George, Beyond Valuations: Long-Term Financial Risks from the AI Build-Out. Federal Reserve Bank of Kansas City, Economic Bulletin, 2026-10-05. Charts 1, 3 and 4 and endnotes. Accessed 2026-10-18.

74. Axios issuance report. Madison Mills, 2026-10-16, citing FT. Accessed 2026-10-18.

75. Rising borrowing costs signal funding risks for global AI expansion: Jefferies report. Economic Times/ANI, 2026-10-17. Full HTML retrieved directly; original Jefferies note not retrieved. Accessed 2026-10-18.

76. Brace for Impact: Protecting People and the Economy when the AI Bubble Bursts. Open Markets Institute, 2026-10-16. Advocacy report; not a holdings census. Accessed 2026-10-18.

77. Microsoft FY2024 annual report, Note 1. Historical useful-life policy, effective FY2023. Accessed 2026-10-18.

78. Meta 2025 10-K, Note 1. 2026-01-29; source 22 rechecked for hardware/building lives on 2026-10-18.

79. Independent Claude debt-markets cross-check. Claude (Opus 5.5), 2026-10-18; edits 1–13 applied 2026-10-18 within the note/AI debt CSV scope; circular CSV reverse links not edited.

80. Meta, BlackRock El Paso venture announcement. 2026-07-28. Directly fetched 2026-10-18; announced financing/conditional closing, not verified outstanding debt.

81. Meta Q2 10-Q, Notes 5, 8 and 13. Source 3 directly re-fetched 2026-10-18: Hyperion lease/RVG, debt fair value and conditional El Paso RVG.

82. Alphabet Q2 10-Q, debt and equity disclosures. Source 8 directly re-fetched 2026-10-18: notes fair value, broader principal perimeter and $40B ATM authorization.

83. Microsoft FY2026 10-K, Notes 1, 10, 13 and contractual obligations. Source 16 directly re-fetched 2026-10-18: face/fair values, useful lives and lease payments.

84. Oracle FY2026 10-K and August 10-Q. Sources 20 and 67 directly re-fetched 2026-10-18: principal components and August debt fair value. All five SEC fetches used generic User-Agent research-bot contact@example.com.

85. Cohen, Greg, Cooper Killen and Simon Lau, Tail Risk for Banks Posed by Investments in Generative Artificial Intelligence. Chicago Fed Insights, February 2026. Directly fetched 2026-10-18; FR Y-14Q large-bank analysis and separately attributed MSCI data-centre lending.

86. IMF, Global Financial Stability Report, Chapter 1: Beneath the Resilience: Rising Risks, Unresolved Vulnerabilities. October 2026. PDF directly fetched and text checked 2026-10-18; broad BDC/insurer/pension findings, not AI-specific holdings.

87. Coinpaper, AI Debt Issuance Falls From $113B to $23B While Oracle Loans Trade Below Par. 2026-10-11. Directly fetched 2026-10-18; tertiary relay of FT/Morgan Stanley, original bank dataset not retrieved.

88. CoreWeave DDTL 4.0 and 5.5 releases. Sources 32 and 33 directly re-fetched 2026-10-18; non-recourse/collateral/insurance participation and prior $3.1B DDTL 5.0.

89. CoreWeave, DDTL 3.0 first amendment, 8-K Item 1.01. 2026-01-02. Directly fetched 2026-10-18; October 28 changes cure limits, not automatic default.

90. Cruz, Stephanie, Meta Pays 7.5% on $12.5 Billion Data Centre Debt to Keep Its AI Dream Alive. IBTimes UK, 2026-07-29. Accessed 2026-10-18; reported pricing/2048/S&P rating, C; ~0.4pp comparison and expected Fitch rating relayed in Claude review, not independently certified.

91. CNBC, CoreWeave/Oracle AI credit stress. Published 2026-10-19, updated 2026-10-20; full AMP body re-fetched 2026-10-21, explicitly Monday observations.

92. CNBC, CoreWeave debt and Janus Henderson’s Lloyd. 2026-10-20; full AMP body re-fetched 2026-10-21. Market marks are reported snapshots.

93. Financial Times via Financial Post, AI debt worries spark surge in default-insurance trading. 2026-10-20; full syndicated body re-fetched 2026-10-21; not an original bank dataset or synchronized closing tape.

94. Reuters via Boursorama, S&P warns hyperscaler AI liabilities may be greater than known. Marc Jones, 2026-10-19; full Reuters syndication re-fetched 2026-10-21, no rating action.

95. SpaceX, September earnings exhibit. Released 2026-10-20; unaudited balance sheet/income/cash-flow tables directly re-fetched 2026-10-21 using research-bot contact@example.com.

96. SpaceX, June 10-Q, Notes 9 and 17. Quarter ended 2026-06-30; directly re-fetched 2026-10-21 with the same generic User-Agent; names Valor and failed-sale-leaseback accounting.

97. CoreWeave, DDTL 3.0 amendment. December 31, 2025 amendment filed January 2, 2026; re-fetched 2026-10-21 with the same generic User-Agent. Historical contract, not a new October amendment.

98. NMED scheduling order, dismissal notice and discovery notice. Filed 2026-10-16; directly downloaded and text checked 2026-10-21. Older procedural orders newly surfaced, no permit approval.

99. CNBC, October 15 CoreWeave credit stress. Historical 15.78% same-coupon comparator retained from the October 17 verified record and credit CSV; not a new October 21 observation.